Intro page about

How we sampled people, general construction and explanation of Gold-MSI and our genre/gender questions and announing setup of our research

Some graphs with general info about gender, sophistication scores and genre preferences

Under construction

The LCA

Latent Class Analysis or LCA is a psychometric method in which participants are grouped based on how likely they would respond positive to a certain survey item, in our case a song snippet that is either beautiful or not. After running the results of our 119 participants through the LCA, it appeared that only a 3-class model fitted the data appropriately, so that is what you see in the table.

At the top of the table, the currently unnamed classes are visible. The row with class proportions shows how many of our participants where predicted to be in that class, which means that 14% belongs to class 1, 38% to class 2 etc. Below that are the item percentages, which indicate the probability of a person belonging to that class to say that they liked the song, so for instance on Item 8, a person belonging to class 1 has an 8% chance of liking the song, class 2 a 31% chance of liking the song and a person belonging to class 3 a 16% chance of liking the song.

Judging by tables, it appears that the 3 classes can be interpreted as follows: A class that likes very little songs (class 1), a class that likes a lot of the songs (class 2) and a class that lies somewhere in between these 2 classes.

LCA class table

ANOVA of Gold-MSI yields what we suspected

On the right you see the distribution of Gold-MSI scores per class, an ANOVA indicated that there was a difference between classes and post-hoc analysis with bonferonni correction showed this to be only probable for the classes 1 and 2.

Charts of class charateristics

```

Conclusion

Under construction

Discussion

Under construction

Rating the Songs

When starting out our survey, we searched online for known datasets that included musical pieces that were dissected on their musical components. This was important as we wanted to compare musical sophistication with musical pieces and we needed information about the structure of these pieces. Since this yielded no results, we decided to each supply 5 instrumental songs to a playlist on Spotify. These songs needed to be instrumental to control for the influence of language on the perception of beauty. After we compiled 30 songs, we then used 3 musical experts with more than 10 years of formal training to rate them on 9 components on a 10-point Likert-scale, copying the method used in the article of Aljanaki et al. (2016). The following components were:

• Tempo: the general pulse of the song, ranging from very slow (1) to very fast (10)

• Articulation: The rhythmic articulation of each song, ranging from very staccato (1) to completely legato (10), staccato are separate notes with rests in between, legato notes are notes that are strung together.

• Mode: overall mode and feel of the songs, ranging from minor (1) to major (10)

• Intensity: overall loudness and crescendos and decrescendos in a song, ranging from 1 (pianissimo) to 10 (fortissimo)

• Tonalness: overall tonalness of the composition, ranging from (1) atonal, with no discernable mode or key to tonal (10) with no use of “outside” extensions and very clear discernable key and mode

• Pitch: overall distribution of the pitches, ranging from all bass (1) to all treble (10)

• Melody: overall presence and dominance of melody, ranging from very unmelodious (1) to very melodious (10)

• Rhythmic Clarity: overall presence of a pulse, ranging from very vague (1) to very firm (10)

• Rhythmic Complexity: the extent to which different meters, odd tempo’s or complex rhytmic patterns are utilized, ranging from very simple (1) to very complex (10)

After all songs were rated, we selected 15 songs to include on our survey based on A) Feature Representability and B) Reliability

---

A) Feature Representability
The panel on the right is interactive, hover over a point with your mouse to find out more

The combined box and jitterplot shows the overall distribution of the characteristics of the selected songs. The boxplot represents the feature values of all 30 songs. The jitterplot shows the feature values of the 15 songs we selected for our survey.

Examining the jitterplot, it becomes apparent that our selection covers quite a large range for most components, with a range of around 6 for most components. Certain interest should be given towards the component of Pitch, which features mostly average Bass/Treble compositions, with 1 lower range song.

Overall this looks to be an okay distribution of songs, given that the playlist was compiled by 6 different people with different preferences. For some components however, a more extreme rating would be preferred so we would’ve had more room to examine the eventual class differences.

Box and jitterplot of average expert rating per song

Assessing reliability

To start our selection of 15 songs, we first estimated the reliability of the expert ratings per song. To do this we computed distance scores between each of the 3 experts. For example, each rater provided a rating of the component Tempo for a given song. The first rater gave it a 5, the second rater gave it a 6 and the third a 7. The distance could then be calculated by taking the distance between the first and the second rater (6 - 5 = 1), the distance between the second and the third rater (7 - 6 = 1) and the distance between the first and the third rater (7 - 5 = 2). We then summed the difference (1 + 1 + 2 = 4), which provided an estimate of rater consensus on the component tempo.

Subsequently, this was done for all components per song, and then all the reliability scores per component were summed to give an estimate of overall reliability. The table on the right shows these scores for all 30 songs.

As can be seen from the table, the reliability scores range between 26 and 64, with a lower score representing better consensus on that song. Based on these scores, we estimated a cut-off point for song selection (reliability score < 45), and used this to select our songs. Upon examining our prior selection however, it became apparent that the distribution of tempo ratings was skewed to favour higher tempos and not enough atonal songs (low score on tonalness). To correct for this we decided to swap the song Sesiu Nata Drama (reliability score of 46) of the song The Kiss (reliability score of 42), to make sure our songs represented most of the component ranges. In the next segment we will examine this further.

Reliability scores per song

Snippets

Blueming

Bygone Bumps

Cia Pat

Decision (Price of Love)

Elysium

Firth Of Fifth

Less Is Moi

Married Life

Resolver

Scarface Theme

Single Petal Of A Rose

Song For A New Beginning

syro u473t8+e

Šešių Natų Drama _ Drama In Six Notes

USA III Rail

Full songs

Contact information and Names

Hello, and welcome to our portfolio. We are students from the honours course The Data Science of Everyday Music Listening coordinated by dr. J.A. Burgoyne, and we wanted to know more about the beauty of music. In our brainstorm sessions we concluded that the experience of musical beauty differs from person to person. We wanted to know if someone’s musical sophistication influenced what songs they deemed beautiful. In this portfolio you will find the method and results of our research and we hope you will enjoy it. Sincerely, Willem Pleiter, Kristina Savickaja, Xiaoqing Li, Denise Quek, Nikita van ‘t Rood and Esther Liefting.

For further information, please contact us at estherliefting@student.uva.nl